DeepSeek V4.1 Flash Launch Offers Breakthrough Speed and Lower Cost

DeepSeek V4‘s newest member, DeepSeek-V4.1-Flash, launched on Thursday as the smallest model in a new architectural family built to run faster and cheaper than its predecessors.

The Chinese AI startup had been quietly testing the model through a limited-time beta that began a day earlier, according to TechNode, which described it as an interim release using a new architecture with native multimodal support.

Reuters separately reported on the launch the same day. Independent benchmark testing recorded throughput near 427 tokens per second in one run, positioning the new model as a speed-focused release rather than a pure capability jump over DeepSeek’s larger Pro-tier offerings.

What DeepSeek V4 Flash Changes For Pricing And Traffic

DeepSeek is routing a portion of its Pro-tier traffic to the new Flash architecture, according to reporting from Intelligent Living, which also flagged lower Flash-tier pricing tied to the launch.

The earlier Flash variant already ran as an efficiency-focused mixture-of-experts model with 284 billion total parameters and 13 billion active per token, alongside a 1 million-token context window, per listings on DeepInfra. The V4.1 release appears to extend that efficiency approach rather than replace it outright.

What that means in practice, rate limits, regional availability, and which subscription tiers can access the new architecture, has not been formally specified by the company.

Benchmark Claims Still Unsettled

Community testing shared on Reddit claimed V4.1 Flash reached 98% of a rival model’s benchmark score while using a fraction of the compute, and separate comparisons suggested it ran up to 38% faster with 43% fewer tokens than an earlier Flash vision variant.

None of these figures come from DeepSeek’s own materials, and the company has not published formal benchmark tables alongside the release. That gap between company messaging and third-party testing leaves open how the model actually stacks up against Pro-tier alternatives on harder reasoning tasks, even as its speed and pricing advantages look real on ordinary workloads.

Read Next: Inside the AI Model Race Reshaping Compute Costs